New energy photovoltaic module defect classification method based on information entropy cognitive neural field

Abstract

This invention discloses a defect classification method for new energy photovoltaic modules based on information entropy cognitive neural fields, belonging to the field of defect classification technology. The method includes the following steps: S1, receiving the original electroluminescent image of the new energy photovoltaic module and encoding it to obtain a high-dimensional dense feature vector; S2, reconstructing the high-dimensional dense feature vector to obtain a reconstructed core defect feature vector; S3, receiving the reconstructed core defect feature vector and strengthening it through a neural field lateral interaction differential equation; S4, adjusting the dynamic representation unit of photovoltaic defect features and using IT layer neurons to output the photovoltaic module defect classification result label. This invention not only structurally addresses the shortcomings of existing visual quality inspection models that lack an adaptive attention mechanism, but also ensures that the system can still provide a unique and high-confidence quality classification conclusion in complex and non-standard photovoltaic production environments.

Description

Technical Field

[0001] This invention belongs to the field of defect classification technology, specifically relating to a defect classification method for new energy photovoltaic modules based on information entropy cognitive neural field. Background Technology

[0002] With the rapid iteration of the global new energy industry, the manufacturing quality control of photovoltaic modules (solar panels) has become a core factor determining the power generation efficiency and lifespan of photovoltaic power plants. Electroluminescence (EL) imaging technology is widely used on automated photovoltaic production lines to detect internal defects in the modules.

[0003] However, real-world industrial quality inspection scenarios face severe challenges of "small sample size" and "strong interference": on the one hand, the frequency of fatal micro-defects that lead to component scrap (such as microcracks, micro-scratches, broken gates, etc.) is extremely low, making it difficult to collect a large number of samples; on the other hand, these micro-defects are often masked by complex polycrystalline silicon grain boundary textures and uneven brightness backgrounds.

[0004] Traditional deep learning networks (such as convolutional neural networks and visual Transformers) heavily rely on massive amounts of labeled data when handling such tasks. Under small sample conditions, traditional networks are prone to overfitting due to their large number of parameters, resulting in persistently high false negative and false positive rates in actual production lines.

[0005] To overcome the limitations of traditional deep learning in small-sample scenarios, those skilled in the art have attempted to introduce hierarchical cognitive neural models based on the brain's visual cognitive mechanisms (such as dynamic neural field models simulating the V1-V2-V4-IT visual pathway). These models typically utilize pre-trained visual networks to extract high-dimensional features from images and input them into dynamic neural field equations simulating the higher visual cortex (V4 area), achieving classification through the "lateral inhibition and excitation" mechanism between neurons.

[0006] However, existing cognitive neural field technology reveals the following serious inherent technical defects when applied to photovoltaic EL image defect detection:

[0007] 1. Direct input of dense, high-dimensional features triggers the "curse of dimensionality," failing to meet the real-time requirements of production lines: Existing neural field models, after extracting features using pre-trained networks (such as ViT), typically use hundreds or thousands of dense, high-dimensional features directly as input to the neural field dynamics equations. When solving the differential equations, the Difference of Gaussians (DoG) kernel function needs to perform complex distance calculations in this high-dimensional space. This approach leads to an exponential explosion in computational load, causing a severe "curse of dimensionality" and resulting in extremely long system inference times, making it impossible to meet the "millisecond-level" real-time quality inspection cycle requirements of new energy production lines.

[0008] 2. Lack of an adaptive filtering mechanism based on biomimetic vision, core micro-defects are easily drowned out by background noise: In photovoltaic EL images, over 90% of the pixels are normal silicon wafer grain boundaries and background textures. Existing cognitive models lack evaluation and dimensionality reduction filtering mechanisms specific to features, using an "equal weight" approach to process all input dimensions. This results in a massive amount of normal polycrystalline silicon background texture acting as "redundant noise," severely diluting the feature weights that truly characterize micro-defects such as hidden cracks and scratches. In the subsequent neural field lateral inhibition interaction, background noise generates strong interference, causing the model's classification accuracy for photovoltaic micro-defects to be extremely unstable under limited sample conditions, making it difficult to meet industrial-grade quality inspection standards.

[0009] In summary, designing an efficient and adaptive feature selection and dimensionality reduction mechanism before high-dimensional features of photovoltaic images enter dynamic neural field inference, accurately removing redundant polycrystalline silicon background noise and locking in the most discriminative defect features, is a core technical challenge that urgently needs to be overcome in the field of intelligent quality inspection of new energy photovoltaics.

[0010] Existing technology discloses a hierarchical cognitive neural model (HCNM) for few-sample image classification, which simulates the ventral visual pathway (V1-V2-V4-IT) in human visual cognition. The specific implementation scheme is as follows:

[0011] Feature extraction module (simulating V1 / V2 regions): A pre-trained Visual Transformer (ViT) is used as the feature extractor with frozen weights. The input image is encoded into a sequence of high-dimensional feature vectors for subsequent representation.

[0012] Feature representation neural field (simulating V4 area): Discrete dynamic neural field equations are used to describe neuronal activity, simulating the representation of visual features in the V4 area of ​​the visual cortex. The intensity of lateral interactions between neurons is determined by the difference of Gaussians (DoG) kernel function.

[0013] Category representation neural field (simulated IT region): IT neurons are built to store category information, and their activation state is determined by the feedforward input of the V4 layer and the lateral interaction within the IT layer.

[0014] Representative Neuron (RPN) selection algorithm: To improve efficiency, key representative neurons are identified by clustering in the feature space, reducing the computational scale during prediction.

[0015] Scale-Adaptive Adjustment Algorithm: During the prediction phase, the system dynamically adjusts the lateral interaction scale of the V4 layer based on the activation state of the IT layer using the scale-adaptive algorithm (Algorithm2). This continues until only one class of neurons is activated.

[0016] When the above-mentioned existing technology is applied to the detection of defects in electroluminescence (EL) images of new energy photovoltaic modules, the following significant technical defects are exposed:

[0017] Redundancy in feature dimensions leads to heavy computational burden: The model directly uses the dense high-dimensional features extracted by ViT as the neural field input, which results in excessive consumption of computational resources and storage space when solving the neural field dynamic equations (especially the DoG kernel function calculation) in the high-dimensional feature space, making it difficult to meet the real-time detection requirements of photovoltaic production lines.

[0018] The lack of a feature selection mechanism leads to weak anti-interference ability: Existing technologies do not specifically screen or reduce the dimensionality of extracted high-dimensional image features. In photovoltaic EL images, complex grain boundary background textures and minute defect features (such as microcracks and scratches) coexist. Existing technologies cannot adaptively "focus" on defect features, resulting in a large amount of redundant background noise interfering with the lateral inhibition mechanism of the neural field, which limits the accuracy of defect classification under small sample conditions. Summary of the Invention

[0019] To address the problems of existing hierarchical cognitive neural models having huge computational overhead when processing photovoltaic module images, being susceptible to polycrystalline silicon background noise, and failing to meet the needs of real-time industrial detection, this invention proposes a new energy photovoltaic module defect classification method based on information entropy cognitive neural field.

[0020] The technical solution of this invention is: a defect classification method for new energy photovoltaic modules based on information entropy cognitive neural field, comprising the following steps:

[0021] S1. Receive the original electroluminescent image of the new energy photovoltaic module, encode it using the visual feature preliminary coding unit, and obtain a high-dimensional dense feature vector.

[0022] S2. The high-dimensional dense feature vector is reconstructed using the photovoltaic defect attention reconstruction unit to obtain the reconstructed core defect feature vector.

[0023] S3. The photovoltaic defect feature dynamic characterization unit receives the reconstructed core defect feature vector and performs enhancement processing through the neural field lateral interaction differential equation.

[0024] S4. The photovoltaic defect feature dynamic characterization unit is adjusted by using the photovoltaic defect category adaptive decision unit, and the photovoltaic module defect classification result label is output by the IT layer neuron.

[0025] Furthermore, high-dimensional dense feature vectors The expression is:

[0026] ;

[0027] in, This represents the initial input feature sequence after image slicing. This indicates a bullish self-attention strategy. This indicates the operation of a multilayer sensor. Presentation layer normalization operation.

[0028] Furthermore, S2 includes the following sub-steps:

[0029] S21. Calculate the overall information entropy of high-dimensional dense feature vectors;

[0030] S22. Calculate the attention weight based on the overall information entropy of the high-dimensional dense feature vector;

[0031] S23. Construct a mask operator based on the attention weight;

[0032] S24. Calculate the reconstructed core defect feature vector based on the mask operator.

[0033] Furthermore, attention weight The expression is:

[0034] ;

[0035] in, The overall information entropy represents the high-dimensional dense feature vector. This represents the average intra-class information entropy.

[0036] Furthermore, in S23, if the attention weight exceeds the threshold, the mask operator is 1; otherwise, the mask operator is 0.

[0037] Furthermore, the reconstructed core defect feature vector The expression is:

[0038] ;

[0039] in, This represents the masking operator. This represents the initial high-dimensional dense feature vector of the photovoltaic module extracted by the primary visual feature encoding layer. This represents the Hadama product.

[0040] Furthermore, the expression for the lateral interaction differential equation of the neural field is:

[0041] ;

[0042] ;

[0043] in, Indicates the first The rate of evolution of the activation state of a neuron over time Represents a nonlinear response function with boundaries. This represents the total number of neurons involved in photovoltaic feature representation in the V4 dynamic neural field. Indicates a side-to-side interactive kernel. Indicates input up to the number The photovoltaic core defect feature vector of each neuron Indicates input up to the number The photovoltaic core defect feature vector of each neighboring neuron Indicates the first One neuron in Activation value at time, Indicates the first The basic activation function output of each neuron This represents the activation value of a neuron. This represents the resting potential parameter of a neuron. express Time of the first The external photovoltaic feature stimulus input received by each neuron This represents the amplitude weights of the local excitation effect in the Gaussian difference kernel function. The amplitude weights represent the global suppression effect in the Gaussian difference kernel function. Represents an exponential function. This represents the lateral interaction scale of the neural field. This represents the transverse inhibition scale parameter of the neural field. Represents the cosine distance.

[0044] Furthermore, the activation state of neurons in the IT layer The expression is:

[0045]

[0046] in, Represents a nonlinear response function with boundaries. This represents the total number of neurons in the IT category decision neural field that represent different photovoltaic module defect categories. Indicates the first A feature vector representing photovoltaic defect categories, represented by individual IT neurons. Indicates the first A feature vector representing the photovoltaic defect category of each neighboring IT neuron. This represents the weight of lateral interactions between different types of neurons within the IT neural field. Indicates the first The activation state values ​​of the neighboring IT neurons at time t. Indicates the first The output of the basic activation function of each neighboring IT neuron Indicates the first The activation state value of each IT neuron. This represents the activation value of a neuron. This indicates the first one in the preceding V4 neural field. The basic activation function output of each neuron Indicates the first [unclear] from the V4 neural field The first neuron to the IT neural field The weights of feedforward synaptic connections between individual IT neurons This represents the resting potential parameter of an IT neuron.

[0047] Furthermore, the photovoltaic module defect classification result labels The expression is:

[0048] ;

[0049] in, Indicates the first The activation state value of each IT neuron. This indicates the moment when the category decision neural field reaches a stable state after dynamic evolution.

[0050] The beneficial effects of this invention are:

[0051] (1) This invention breaks through the bottleneck of traditional cognitive models directly calculating high-dimensional dense features, and innovatively introduces an information entropy defect attention reconstruction unit between the encoding layer and the neural field representation layer. By constructing a mask operator based on information entropy weights, the initial high-dimensional features are spatially reduced to reconstruct a highly sparse core defect feature vector. From a mathematical perspective, this improvement greatly reduces the distance calculation and iteration dimension of the Gaussian difference kernel function in the V4 region lateral interaction differential equation, effectively avoiding the dimensionality curse caused by high-dimensional operations, thereby enabling a qualitative leap in the system's inference speed, perfectly matching the "millisecond-level" real-time defect detection cycle of the new energy photovoltaic automated production line.

[0052] (2) In small-sample photovoltaic image detection, this invention innovatively adopts the ratio of overall information entropy to average intra-class information entropy as a quantitative standard for evaluating the specificity of defect features. This mathematical mechanism can automatically identify and eliminate normal polycrystalline silicon grain boundary textures that have small differences between different photovoltaic modules but large fluctuations within normal regions (i.e., low discrimination power) as redundant noise. Through this rigorous feature purification process, it is ensured that the lateral inhibition and excitation of the neural field are entirely driven by the most valuable micro-defects such as hidden cracks and scratches. This effectively overcomes the overfitting phenomenon of traditional deep networks with very few defect samples, significantly reduces the false negative and false positive rates on actual production lines, and significantly improves the accuracy and robustness of classification.

[0053] (3) This invention not only simulates the ventral visual pathway of the human brain in its macroscopic architecture, but also endows the industrial quality inspection system with the biomimetic ability of selective attention in its microscopic logic. Through the adaptive focusing extraction of micro-defect features by the front-end information entropy module, and the dynamic gradient feedback adjustment of the V4 lateral interaction scale by the back-end IT neural field based on the classification activation state, the system can automatically adjust the interactive receptive field for repeated verification when encountering suspected hidden crack defects with extremely low contrast and blurred edges. This not only structurally makes up for the shortcomings of existing visual quality inspection models that lack adaptive attention mechanisms, but also ensures that the system can still give a unique and high-confidence quality classification conclusion in complex and non-standard photovoltaic production environments. Attached Figure Description

[0054] Figure 1 This is a flowchart of a new energy photovoltaic module defect classification method based on information entropy cognitive neural field. Detailed Implementation

[0055] The embodiments of the present invention will be further described below with reference to the accompanying drawings.

[0056] like Figure 1 As shown, this invention provides a defect classification method for new energy photovoltaic modules based on information entropy cognitive neural fields, including the following steps:

[0057] S1. Receive the original electroluminescent image of the new energy photovoltaic module, encode it using the visual feature preliminary coding unit, and obtain a high-dimensional dense feature vector.

[0058] S2. The high-dimensional dense feature vector is reconstructed using the photovoltaic defect attention reconstruction unit to obtain the reconstructed core defect feature vector.

[0059] S3. The photovoltaic defect feature dynamic characterization unit receives the reconstructed core defect feature vector and performs enhancement processing through the neural field lateral interaction differential equation.

[0060] S4. The photovoltaic defect feature dynamic characterization unit is adjusted by using the photovoltaic defect category adaptive decision unit, and the photovoltaic module defect classification result label is output by the IT layer neuron.

[0061] The specific technical problem to be solved by this invention (the purpose of the invention) includes:

[0062] 1. Addressing the "curse of dimensionality" caused by high-dimensional features and improving the real-time performance of photovoltaic production line inspection: Existing technologies directly input dense features extracted from pre-trained networks into a dynamic neural field, resulting in excessively high dimensionality in solving the dynamic differential equations of the V4 and IT regions and lagging system response. This invention aims to introduce an information entropy feature attention evaluation module to achieve adaptive spatial dimensionality reduction and feature reconstruction before features enter the neural field, thereby significantly reducing computational load and ensuring the system meets the millisecond-level quality inspection cycle requirements of new energy production lines.

[0063] 2. Addressing the problem of complex background interference and improving the detection accuracy of minute defects in small sample scenarios: Existing technologies lack targeted feature screening mechanisms, leading to the normal grain boundary background texture of photovoltaic silicon wafers acting as redundant noise that interferes with the lateral suppression judgment of the neural field. This invention aims to utilize a weighted allocation mechanism of overall information entropy and intra-class information entropy to adaptively eliminate low-specificity background interference, accurately locate and enhance the core visual features of minute defects such as hidden cracks and scratches, significantly improving the defect recognition accuracy of the system under conditions with very few samples.

[0064] 3. A biomimetic industrial inspection solution with an adaptive attention mechanism is provided: Addressing the technical shortcomings of existing industrial inspection models, such as the lack of a biomimetic attention mechanism and weak robustness, this invention constructs a closed-loop cognitive architecture from feature purification to scale adaptive adjustment by simulating the "selective attention" logic of the human brain's visual system. This aims to improve the anti-interference capability and decision interpretability of classification models in complex photovoltaic production environments.

[0065] Visual Feature Preliminary Encoding Unit: This unit contains a pre-trained Visual Transformer (ViT) network model, whose function is to receive the original electroluminescence image of the new energy photovoltaic module to be detected. Through image slicing processing and a self-attention mechanism, this unit transforms the complex original pixel signal into a high-dimensional dense feature vector representing the basic texture, edges, and light intensity distribution of the photovoltaic module.

[0066] The photovoltaic defect attention reconstruction unit based on information entropy: This unit, located between the encoding unit and the subsequent neural field, is the core feature filtering layer of this invention. It includes an adaptive discretization component and a feature specificity evaluation component. This unit adaptively identifies and removes redundant features such as normal grain boundary background textures in photovoltaic silicon wafers by calculating the overall information entropy and intra-class information entropy of features in each dimension across different photovoltaic defect categories. Finally, this unit reconstructs a core visual feature vector in the reduced-dimensional space that contains only key defect information such as microcracks, scratches, and broken grids.

[0067] The photovoltaic defect feature dynamic characterization unit comprises a discrete dynamic network consisting of multiple neuron nodes. It receives the reconstructed core defect feature vector and uses the differential equations of lateral interaction (local excitation and global inhibition) of the neural field to perform feature dynamic characterization within the purified core feature space. This unit simulates the signal enhancement mechanism of the visual cortex, amplifying minute and ambiguous hidden crack signals during dynamic evolution, thereby improving the system's ability to identify fine-grained defects.

[0068] Photovoltaic Defect Category Adaptive Decision Unit: This unit contains high-level decision neurons with specific defect category preferences (such as normal modules, microcracked modules, scratched modules, etc.). This unit is coupled to the V4 unit through a simulated synaptic connection mechanism. An integrated closed-loop adaptive feedback regulator can adjust the lateral interaction scale parameters of the V4 layer in real time based on the activation response of the IT layer neurons. Finally, the system outputs the corresponding photovoltaic module defect classification result label by determining the IT neuron with the highest activation value.

[0069] In this embodiment of the invention, high-dimensional dense feature vectors The expression is:

[0070] ;

[0071] in, This represents the initial input feature sequence after image slicing. This indicates a bullish self-attention strategy. This indicates the operation of a multilayer sensor. Presentation layer normalization operation.

[0072] Multi-head self-attention operation is used to calculate the global dependency and feature correlation between features in different regions of photovoltaic images; multilayer perceptron operation is used to perform nonlinear mapping and dimensional transformation on the extracted features; layer normalization is used to standardize the data of the intermediate layers of the network in terms of feature dimensions to stabilize the feature distribution within the model.

[0073] The primary encoding layer for photovoltaic visual features is responsible for spatial location encoding and local feature extraction of the original photovoltaic electroluminescence image signal, initially capturing silicon wafer texture, brightness distribution, and potential defect contours. This layer consists of a pre-trained visual Transformer module, the core of which is the self-attention matrix.

[0074] The input photovoltaic defect image tensor is processed through image slice embedding to obtain a sequence. The output of this module is a feature vector. The photovoltaic features extracted by this module often contain a massive amount of normal polycrystalline silicon background texture, exhibiting high dimensionality and strong redundancy, and are used as input for subsequent attention mechanisms.

[0075] In this embodiment of the invention, S2 includes the following sub-steps:

[0076] S21. Calculate the overall information entropy of high-dimensional dense feature vectors;

[0077] S22. Calculate the attention weight based on the overall information entropy of the high-dimensional dense feature vector;

[0078] S23. Construct a mask operator based on the attention weight;

[0079] S24. Calculate the reconstructed core defect feature vector based on the mask operator.

[0080] This study simulates the "selective attention" mechanism of photovoltaic quality inspection experts. By evaluating the inter-class uncertainty (information entropy) and intra-class consistency of feature dimensions, high-dimensional features are significantly enhanced and reduced in dimensionality, filtering out normal grain boundary background noise and retaining only the "core defect visual features" that contribute most to distinguishing microcracks, scratches, etc.

[0081] This module directly determines the computational scale of the subsequent neural field equations, and accelerates the computation of complex photovoltaic images across modes through "background removal and defect purification".

[0082] In this embodiment of the invention, attention weight The expression is:

[0083] ;

[0084] in, The overall information entropy represents the high-dimensional dense feature vector. This represents the average intra-class information entropy.

[0085] In this embodiment of the invention, in S23, if the attention weight exceeds the threshold, the mask operator is 1; otherwise, the mask operator is 0.

[0086] 0 indicates that the feature in this dimension is genuine defect information, while 0 indicates that the dimension is redundant silicon wafer background.

[0087] In this embodiment of the invention, the reconstructed core defect feature vector The expression is:

[0088] ;

[0089] in, This represents the masking operator. This represents the initial high-dimensional dense feature vector of the photovoltaic module extracted by the primary visual feature encoding layer. This represents the Hadama product.

[0090] The Hadamard product multiplies matrices element-wise; its physical meaning is to multiply the matrix by element. The filtering and judgment state (0 or 1) in the process is precisely applied to the high-dimensional feature vector. This achieves the "zeroing" of normal background noise in photovoltaic systems and the lossless preservation of core defect features across each corresponding dimension.

[0091] In this embodiment of the invention, the expression for the lateral interaction differential equation of the neural field is:

[0092] ;

[0093] ;

[0094] in, Indicates the first The rate of evolution of the activation state of a neuron over time Represents a nonlinear response function with boundaries. This represents the total number of neurons involved in photovoltaic feature representation in the V4 dynamic neural field. Indicates a side-to-side interactive kernel. Indicates input up to the number The photovoltaic core defect feature vector of each neuron Indicates input up to the number The photovoltaic core defect feature vector of each neighboring neuron Indicates the first One neuron in Activation value at time, Indicates the first The basic activation function output of each neuron This represents the activation value of a neuron. This represents the resting potential parameter of a neuron. express Time of the first The external photovoltaic feature stimulus input received by each neuron This represents the amplitude weights of the local excitation effect in the Gaussian difference kernel function. The amplitude weights represent the global suppression effect in the Gaussian difference kernel function. Represents an exponential function. This represents the lateral interaction scale of the neural field. This represents the transverse inhibition scale parameter of the neural field. Represents the cosine distance.

[0095] The high-level feature representation neural field layer for photovoltaic defects is used to characterize the topological features of photovoltaic defects in the reconstructed low-dimensional space. Each neuron represents a point in the defect feature space, and the feature distribution of a specific defect is formed through "lateral interactions" between neurons.

[0096] In this embodiment of the invention, the activation state of neurons in the IT layer The expression is:

[0097]

[0098] in, Represents a nonlinear response function with boundaries. This represents the total number of neurons in the IT category decision neural field that represent different photovoltaic module defect categories. Indicates the first A feature vector representing photovoltaic defect categories, represented by individual IT neurons. Indicates the first A feature vector representing the photovoltaic defect category of each neighboring IT neuron. This represents the weight of lateral interactions between different types of neurons within the IT neural field. Indicates the first The activation state values ​​of the neighboring IT neurons at time t. Indicates the first The output of the basic activation function of each neighboring IT neuron Indicates the first The activation state value of each IT neuron. This represents the activation value of a neuron. This indicates the first one in the preceding V4 neural field. The basic activation function output of each neuron Indicates the first [unclear] from the V4 neural field The first neuron to the IT neural field The weights of feedforward synaptic connections between individual IT neurons This represents the resting potential parameter of an IT neuron.

[0099] In this embodiment of the invention, the photovoltaic module defect classification result label The expression is:

[0100] ;

[0101] in, Indicates the first The activation state value of each IT neuron. This indicates the moment when the category decision neural field reaches a stable state after dynamic evolution.

[0102] The photovoltaic defect category decision neural field layer stores abstract photovoltaic defect category labels (such as normal, microcrack, broken grid, scratch, etc.) and performs classification and discrimination. IT neurons and V4 neurons are connected in a feedforward manner through a synaptic weight matrix.

[0103] The system includes a closed-loop feedback controller, if If the value is not equal to 1 (i.e., the classification is fuzzy or there is no activation, and the defect type cannot be determined), then the step size is adjusted according to the gradient. Dynamically adjust the interaction scale of the V4 layer :

[0104] ;

[0105] in, This represents the lower bound threshold of the dynamic adjustment of the transverse interaction scale of the V4 neural field (i.e., the minimum permissible interaction receptive field). This represents the upper bound threshold for the dynamic adjustment of the transverse interaction scale of the V4 neural field. This represents the new (updated) V4 neural field lateral interaction scale parameters output after closed-loop feedback calculation.

[0106] This mechanism ensures that the system can still provide a unique and definitive quality inspection conclusion even in a small sample environment with micro-defects in photovoltaics.

[0107] Those skilled in the art will recognize that the embodiments described herein are intended to help the reader understand the principles of the invention, and should be understood that the scope of protection of the invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific modifications and combinations based on the technical teachings disclosed in this invention without departing from the spirit of the invention, and these modifications and combinations are still within the scope of protection of this invention.

Claims

1. A defect classification method for new energy photovoltaic modules based on information entropy cognitive neural fields, characterized in that, Includes the following steps: S1. Receive the original electroluminescent image of the new energy photovoltaic module, encode it using the visual feature preliminary coding unit, and obtain a high-dimensional dense feature vector. S2. The high-dimensional dense feature vector is reconstructed using the photovoltaic defect attention reconstruction unit to obtain the reconstructed core defect feature vector. S3. The photovoltaic defect feature dynamic characterization unit receives the reconstructed core defect feature vector and performs enhancement processing through the neural field lateral interaction differential equation. S4. The photovoltaic defect feature dynamic characterization unit is adjusted by using the photovoltaic defect category adaptive decision unit, and the photovoltaic module defect classification result label is output by the IT layer neuron.

2. The method for classifying defects in new energy photovoltaic modules based on information entropy cognitive neural fields according to claim 1, characterized in that, The high-dimensional dense feature vector The expression is: ； in, This represents the initial input feature sequence after image slicing. This indicates a bullish self-attention strategy. This indicates the operation of a multilayer sensor. Presentation layer normalization operation.

3. The method for classifying defects in new energy photovoltaic modules based on information entropy cognitive neural fields according to claim 1, characterized in that, S2 includes the following sub-steps: S21. Calculate the overall information entropy of high-dimensional dense feature vectors; S22. Calculate the attention weight based on the overall information entropy of the high-dimensional dense feature vector; S23. Construct a mask operator based on the attention weight; S24. Calculate the reconstructed core defect feature vector based on the mask operator.

4. The method for classifying defects in new energy photovoltaic modules based on information entropy cognitive neural fields according to claim 3, characterized in that, The attention weight The expression is: ； in, The overall information entropy represents the high-dimensional dense feature vector. This represents the average intra-class information entropy.

5. The method for classifying defects in new energy photovoltaic modules based on information entropy cognitive neural fields according to claim 3, characterized in that, In step S23, if the attention weight exceeds the threshold, the mask operator is 1; otherwise, the mask operator is 0.

6. The method for classifying defects in new energy photovoltaic modules based on information entropy cognitive neural fields according to claim 3, characterized in that, The reconstructed core defect feature vector The expression is: ； in, This represents the masking operator. This represents the initial high-dimensional dense feature vector of the photovoltaic module extracted by the primary visual feature encoding layer. This represents the Hadama product.

7. The method for classifying defects in new energy photovoltaic modules based on information entropy cognitive neural fields according to claim 1, characterized in that, The expression for the lateral interaction differential equation of the neural field is as follows: ； ； in, Indicates the first The rate of evolution of the activation state of a neuron over time Represents a nonlinear response function with boundaries. This represents the total number of neurons involved in photovoltaic feature representation in the V4 dynamic neural field. Indicates a side-to-side interactive core. Indicates input up to the number The photovoltaic core defect feature vector of each neuron Indicates input up to the number The photovoltaic core defect feature vector of each neighboring neuron Indicates the first One neuron in Activation value at time, Indicates the first The basic activation function output of each neuron This represents the activation value of a neuron. This represents the resting potential parameter of a neuron. express Time of the first The external photovoltaic feature stimulus input received by each neuron This represents the amplitude weights of the local excitation effect in the Gaussian difference kernel function. The amplitude weights represent the global suppression effect in the Gaussian difference kernel function. Represents an exponential function. This represents the lateral interaction scale of the neural field. This represents the transverse inhibition scale parameter of the neural field. It represents the cosine distance.

8. The method for classifying defects in new energy photovoltaic modules based on information entropy cognitive neural fields according to claim 1, characterized in that, The activation state of the neurons in the IT layer The expression is: ； in, Represents a nonlinear response function with boundaries. This represents the total number of neurons in the IT category decision neural field that represent different photovoltaic module defect categories. Indicates the first A feature vector representing photovoltaic defect categories, represented by individual IT neurons. Indicates the first A feature vector representing the photovoltaic defect category of each neighboring IT neuron. This represents the weight of lateral interactions between different types of neurons within the IT neural field. Indicates the first The activation state values ​​of the neighboring IT neurons at time t. Indicates the first The output of the basic activation function of each neighboring IT neuron Indicates the first The activation state value of each IT neuron. This represents the activation value of a neuron. This indicates the first [neuron] in the preceding V4 neural field. The basic activation function output of each neuron Indicates the first [unclear] from the V4 neural field The first neuron to the IT neural field The weights of feedforward synaptic connections between individual IT neurons This represents the resting potential parameter of an IT neuron.

9. The method for classifying defects in new energy photovoltaic modules based on information entropy cognitive neural fields according to claim 1, characterized in that, The photovoltaic module defect classification result label The expression is: ； in, Indicates the first The activation state value of each IT neuron. This indicates the moment when the category decision neural field reaches a stable state after dynamic evolution.

Images